DISSERTATION DEFENSE Author : Zhonghao Liu Advisor : Dr. Jianjun Hu Date : Oct. 29th , 2018 Time : 9:00 am Place : Meeting room 2267, Innovation Center Abstract Machine learning and deep learning methods have been increasingly applied to solve challenging and important bioinformatics problems such as protein structure pre- diction, disease gene identification, and drug discovery. However, the performances of existing machine learning based predictive models are still not satisfactory. The question of how to exploit the specific properties of bioinformatics data and couple them with the unique capabilities of the learning algorithms remains elusive. In this dissertation, we propose advanced machine learning and deep learning algorithms to address two important problems: mislocation-related cancer gene identification and major histocompatibility complex-peptide binding affinity prediction. Our first contribution proposes a kernel-based logistic regression algorithm for identifying potential mislocation-related genes among known cancer genes. Our algorithm takes protein-protein interaction networks, gene expression data, and subcellular location gene ontology data as input, which is particularly lightweight comparing with existing methods. The experiment results demonstrate that our proposed pipeline has a good capability to identify mislocation-related cancer genes. Our second contribution addresses the modeling and prediction of human leukocyte antigen (HLA) peptide binding of human immune system. We present an allele-specific convolutional neural network model with one-hot encoding. With extensive evaluation over the standard IEDB datasets, it is shown that the performance of our model is better than all existing prediction models. To achieve further improvement, we propose a novel pan-specific model on peptide-HLA class I binding affinities prediction, which allows us to exploit all the training samples of different HLA alleles. Our sequence-based pan model is currently the only algorithm not using pseudo sequence encoding — a dominant structure-based encoding method in this area. The benchmark studies show that our method could achieve state-of-the-art performance. Our proposed model could be integrated into existing ensemble methods to improve their overall prediction capabilities on highly diverse MHC alleles. Finally, we present a LSTM-CNN deep learning model with attention mechanism for peptide-HLA class II binding affinities and binding cores prediction. Our model achieved very good performance and outperformed existing methods on half of tested alleles. With the help of attention mechanism, our model could directly output the peptide binding core based on attention weight without any additional post- or pre- processing.

ACM is hosting a 24 hour Code-A-Thon on Friday, October 26th in Swearingen 1D11 at 7PM (and also online) . The Code-A-Thon is open to all majors and all skill levels. If you can’t make the opening event, compete online here: http://www.hackerrank.com/usc-acm-fall-2018-145-division http://www.hackerrank.com/usc-acm-fall-2018-146-division http://www.hackerrank.com/usc-acm-fall-2018-240-division http://www.hackerrank.com/usc-acm-fall-2018-350-division Pick the division of the highest CS course you are enrolled in or have taken (out of 145, 146, 240, and 350). If you haven’t taken any of these CS classes, take the 145 division. If you are a CS graduate student, or have already taken 350, compete in the 350 division. You can pick a lower division, but you won’t be able to compete for prizes. Prizes are:

• 32GB flash drive & wireless keyboard
• Raspberry Pi 3 B+
• 100,000 mAh mobile battery

Let me know if you have any questions, James Coman President, ACM University of South Carolina | Class of 2019 ACM Friendface Page

DISSERTATION DEFENSE Department of Computer Science and Engineering University of South Carolina Author : Chao Chen Advisor : Dr. Gabriel Terejanu Date : Oct. 15th , 2018 Time : 2:30 pm Place : Meeting room 2267, Innovation Center Abstract Normal neural networks trained with gradient descent and back-propagation have received great success in various applications. On one hand, point estimation of the network weights is prone to over-fitting problems and lacks important uncertainty information associated with the estimation. On the other hand, exact Bayesian neural network methods are intractable and non-applicable for real-world applications. To date, approximate methods have been actively under development for Bayesian neural networks, including but not limited to, stochastic variational methods, Monte Carlo dropouts, and expectation propagation. Though these methods are applicable for current large networks, there are limits of these approaches with either under estimation or over-estimation of uncertainty. Extended Kalman filters (EKFs) and unscented Kalman filters (UKFs), which are widely used in data assimilation community, adopt a different perspective of inferring the parameters. Nevertheless, EKFs are incapable of dealing with highly non-linearity, while UKFs are inapplicable for large network architectures. Ensemble Kalman filters (EnKFs) serve as great methodology in atmosphere and oceanology disciplines targeting extremely high-dimensional, non-Gaussian, and nonlinear state-space models. So far, there is little work that applies EnKFs to estimate the parameters of deep neural networks. By considering neural network as a nonlinear function, we augment the network prediction with parameters as new states and adapt the state-space model to update the parameters. In the first work, we describe the ensemble Kalman filter, two proposed algorithms for training both fully-connected and Long Short-term Memory (LSTM) networks, and experiment it with a synthetic dataset, 10 UCI datasets, and a natural language dataset for different regression tasks. To further evaluate the effectiveness of the proposed training scheme, we trained a deep LSTM network with the proposed algorithm, and applied it on five real-world sub-event detection tasks. With a formalization of the sub-event detection task, we develop an outlier detection framework and take advantage of the Bayesian Long Short-term Memory (LSTM) network to capture the important and interesting moments within an event. In the last work, we develop a framework for student knowledge estimation using Bayesian network. By constructing student models with Bayesian network, we can infer the new state of knowledge on each concept given a student. With a novel parameter estimate algorithm, the model can also indicate misconception on each question. Furthermore, we develop a predictive validation metric with expected data likelihood of the student model to evaluate the design of questions.

THESIS DEFENSE Department of Computer Science and Engineering University of South Carolina Author : Lacie Renee Stiffler Advisor : Dr. Bakos Date : October 10th, 2018 Time : 10:00 am Place : 2267, Storey Innovation Center Abstract Artificial neural networks are an effective machine learning technique for a variety of data sets and domains, but exploiting the inherent parallelism in neural networks requires specialized hardware. Typically, computing the output of each neuron requires many multiplications, evaluation of a transcendental activation function, and transfer of its output to a large number of other neurons. These restrictions become more expensive when internal values are represented with increasingly higher data precision. A spiking neural network eliminates the limitations of typical rate-based neural networks by reducing neuron output and synapse weights to one-bit values, eliminating hardware multipliers, and simplifying the activation function. However, a spiking neural network requires a larger number of neurons than what is needed in a comparable rate-based network. In order to determine if the benefits of spiking neural networks outweigh the costs, we designed the smallest spiking neural network and rate-based artificial neural network that achieved 90\% or comparable testing accuracy on the MNIST data set. After estimating the FPGA storage requirements for synapse values of each network, we concluded rate-based neural networks need significantly fewer bits than spiking neural networks.

THESIS DEFENSE Department of Computer Science and Engineering Author : George Akhvlediani Advisor : Dr. Buell Date : September 28th Time : 10:00 am Place : 2265, Storey Innovation Center Abstract This research examines conflicts accompanying the proliferation of computer technology and, more specifically, constellations of dependency in the always expanding volume of software, platforms, and the firms/individuals using them. We identify a pervasive phenomenon of “divarication” in the growing variety of progressively specialized systems and system roles. As software systems enter new thresholds of sophistication, they effectively aggregate many distinct components and protocols. Consequently, we are confronted with a diverse ecology of stratified and thereby incompatible software systems. Software inherits the limitations and potential flaws of its constituent parts, but unlike physical machinery, it isn’t readily disassembled in instances of failure. The individuals using these systems have no means to dissect and analyze their tools, and thus are necessarily dependent on developer assistance. We assert that divarication is a consequence of interfacing, and particularly in the way computer interfaces operate as the sole point of contact between a user and a software system. Taking Charles S. Peirce’s three types of sign (the icon, index, and symbol) into special consideration, we observe that computer interfaces seldom employ iconic representation. In other words, these interfaces do not reflect the interior logic that drives them; they bear no resemblance to their referent(s). Merely “using” software doesn’t promise any insight into how that software works. We argue that this circumstance makes divarication inevitable. Opaque elements are assembled together into opaque wholes, and so the magnitude of this problem will likely scale with increasing software sophistication. As the thesis title indicates, we bring Peirce’s notion of “iconicity” into accompaniment with “interfacing”, forming an abstract paradigm in response to divarication. We intend to infuse a software platform with a recurrent protocol of iconicity, to develop a platform that allows at least partial disassembly and examination of the programs it facilitates. We composed a diagrammatic design scheme; a blueprint for software platforms that might emulate “interfacing iconicity”. We developed a prototype platform, implementing this structural logic. This initial prototype is a rudimentary HTML rendering platform, one that articulates the relationship between plain-text code, its Document Object Model (DOM) representation, and the rendered “page” itself. Currently, this prototype is a useful analog for our argument. Since it offers a distinct perspective on the connections between text markup and its systemic interpretation, it may also have educational utility. However, it is not yet a fully realized implementation of our design paradigm, and at this stage a conclusion on whether the latter genuinely addresses divarication would be premature.

Abstract: Cybersecurity is becoming one of the challenging problems in the connected world because of heterogeneity of networked systems and scale and complexity of cyberspace. Cyber- attacks are not only increasing in terms of numbers but also getting more sophisticated. Cyber- defense for prevention, detection and response to cyber-attacks is an on-going challenge that needs efforts to protect critical infrastructures and private information. Complexity and scale of cyberspace and heterogeneity of networked systems make cybersecurity even more challenging. Almost all organizations are vulnerable to (similar or same) cyber-attacks where information sharing could help prevent future cyber-attacks This talk presents and evaluates an information sharing framework for cybersecurity with the goal of protecting confidential information and networked infrastructures from future cyber- attacks. The proposed framework leverages the blockchain concept where multiple organizations/agencies participate for information sharing (without violating their privacy) to secure and monitor their cyberspaces. This blockchain based framework is to constantly collect high resolution cyber-attack information across organizational boundaries of which the organizations have no specific knowledge or control over any other organizations' data or damage caused by cyber-attacks. Bio: Laurent L. Njilla received his B.S. in Computer Science from the University of Yaoundé 1 in Cameroon, the M.S. in Computer Engineering from the University of Central Florida (UCF) in 2005 and Ph.D. in Electrical Engineering from Florida International University (FIU) in 2015. He joined the Cyber Assurance Branch of the U.S. Air Force Research Laboratory (AFRL), Rome, New York, as a Research Electronics Engineer in 2015. Prior to joining the AFRL, he was a Senior Systems Analyst in the industry sector for more than 10 years. He is responsible for conducting basic research in the areas of hardware design, game theory applied to cyber security and cyber survivability, hardware Security, online social network, cyber threat information sharing, category theory, and blockchain technology. He is the Program Manager for the Cyber Security Center of Excellence (CoE) for the HBCU/MI and the Disruptive Information Technology Program at AFRL/RI. Dr. Njilla’s research has resulted in more than 50 peer-reviewed journal and conference papers and multiple awards including Air Force Notable Achievement Awards, the 2015 FIU World Ahead Graduate award and etc. He is a reviewer of multiple journals and serves on the technical program committees of several international conferences. He is a member of the National Society of Black Engineer (NSBE). Please see http://www.cse.sc.edu/colloquia for further information on colloquia in the Department of Computer Science and Engineering.

You are invited to join the first Women in Computing event of this school year. Pizza will be provided and everyone - all genders and majors is welcome! Topic: Computing Survival Guide When: Tuesday, September 4th, 7 pm to 8 pm Where: Room 2277, IBM Innovation Center/Horizon 2 (the building next to Strom Thurmond Fitness Center that has the IBM logo on the side). Here is a link to this location on Google Maps (https://goo.gl/maps/GwCroqGrfoS2) The main agenda of this first event is to welcome the Class of 2022. The upperclassmen will share advice and are happy to answer questions from the underclassmen about computing, engineering, or USC life. Hope to see you all soon!

DISSERTATION DEFENSE Author : Hongkai Yu Advisor : Dr. Song Wang Abstract Fast and accurate characterization of fiber micro-structures plays a central role for material scientists to analyze physical properties of continuous fiber reinforced composite materials. In materials science, this is usually achieved by continuously cross-sectioning a 3D material sample for a sequence of 2D microscopic images, followed by a fiber detection/tracking algorithm through the obtained image sequence. To speed up this process and be able to handle larger-size material samples, we propose sparse sampling with larger inter-slice distance in cross sectioning and develop a new algorithm that can robustly track large-scale fibers from such a sparsely sampled image sequence. In particular, the problem is formulated as multi-target tracking and Kalman filters are applied to track each fiber along the image sequence. One main challenge in this tracking process is to correctly associate each fiber to its observation given that 1) fiber observations are of large scale, crowded and show very similar appearances in a 2D slice, and 2) there may be a large gap between the predicted location of a fiber and its observation in the sparse sampling. To address this challenge, a novel group-wise association algorithm is developed by leveraging the fact that fibers are implanted in bundles and the fibers in the same bundle are highly correlated through the image sequence. Tracking-by-detection algorithms rely heavily on detection accuracy, especially the recall performance. The state-of-the-art fiber detection algorithms perform well under ideal conditions, but are not accurate where there are local degradations of image quality, due to contaminants on the material surface and/or defocus blur. Convolutional Neural Networks (CNN) could be used for this problem, but would require a large number of manual annotated fibers, which are not available. We propose an unsupervised learning method to accurately detect fibers on the large scale, which is robust against local degradations of image quality. The proposed method does not require manual annotations, but uses fiber shape/size priors and spatio-temporal consistency in tracking to simulate the supervision in the training of the CNN. Due to the significant microscope movement during the data acquisition, the sampled microscopy images might be not well aligned, which increases the difficulties for further large-scale fiber tracking. We design an object tracking system which could accurately track large-scale fibers and simultaneously perform satisfactory image registration. Large-scale fiber tracking task is accomplished by Kalman filters based tracking algorithms. With the assistance of fiber tracking, the registration error is minimized via a physics optimization model embedded with fibers' 3D trajectory constraints. To evaluate the proposed methods, a dataset was collected by Air Force Research Laboratories (AFRL). The material scientists in AFRL used a serial sectioning instrument to cross-section the 3D material samples. During sample preparation, the samples are ground, cleaned, and then imaged. Experimental results on this collected dataset have demonstrated that the proposed frameworks yield significant improvements in large-scale fiber tracking and detection, together with improved image registration. Date : June 25th , 2018 Time : 8:30 am Place : Meeting room 2267, Innovation Center

DISSERTATION DEFENSE Author : Sharaf Malebary Advisor : Dr. Wenyuan Xu Abstract The rapid advancement in wireless technology along with their low cost and easy to deploy have been attracting researchers academically and commercially. Researchers from private and public sectors are investing into enhancing the reliability, robustness, and security of radio frequency (RF) communications to accommodate the demand and enhance lifestyle. RF base communications -by nature- are slower and more exposed to attacks than a wired base (LAN). Deploying such networks in various cutting-edge mobile platforms (e.g. VANet, IoT, Autonomous robots) adds new challenges that impact the quality directly. Moreover, adopting such networks in public outdoor areas make them vulnerable to various attacks (regardless of the attacker motive). Therefore, the quality and security of the communications cannot be neglected especially when developing outdoor wireless applications/networks. While some wireless applications and platforms aim to provide comfort and infotainment, others are more critical to protect and save lives. Thus, the need for mobile broadband connections has been increased to accommodate such applications. The FCC took the first step to regulate and assure the quality when using these technologies by allocating spectrums and issuing standards and amendments (e.g. IEEE802.11a, b, g, n, and p) to deliver reliable and secure communications. In this dissertation, we introduce several problems related to the security and quality of communications in outdoor environments. Although we focus on the ISM-RF bands UHF and SHF (licensed and unlicensed) and their applications nevertheless, the concept of propagating signals through the air for communications remain the same across other bands. Therefore, problems and solutions in this work can be adopted and applied to different wireless technologies with respect to environment and mobility. Date : June 13th , 2018 Time : 10:00 am Place : Meeting room 2267, Innovation Center